Grenoble H05 independent review protocol

# A10 Grenoble H05 — independent scientific review protocol v0

**Status:** RESEARCH / REVIEW PREPARATION — NO VERDICT  
**Lane:** S — scientific evidence  
**Queue action:** #44  
**Target artifact:** forthcoming A10 Grenoble H05 RESEARCH / PROVISIONAL dry-run

## Purpose and boundary

This protocol lets an independent reviewer assess the A10 output as soon as Lane D produces it. It tests whether the dry-run is traceable, reproducible, scientifically interpretable at its claimed evidence status, and appropriately fail-closed.

This document does **not** calculate A10, select Lane D implementation choices, infer provider semantics, validate H05, promote evidence, or issue a verdict on an artifact that does not yet exist. A merge of A10 research work would not by itself resolve any H05 HOLD.

## Required review bundle

The reviewer must receive an immutable or content-addressed bundle containing:

1. the result artifact and machine-readable QA/provenance output;
2. commit SHA and exact command/configuration used;
3. source manifests, retrieval timestamps, stable identifiers, byte sizes and SHA-256 checksums;
4. Grenoble FUA boundary identity/version and geometry checksum;
5. EEA PM2.5 artifact identity and checksum;
6. GHSL product, epoch, resolution, tile identities and checksums;
7. runtime and dependency versions;
8. logs for coverage, exclusion, alignment and sensitivity checks;
9. explicit records for every provisional assumption and every downstream output that depends on it.

A mutable URL, filename, chart, or prose summary alone is insufficient evidence.

## Review procedure

Review every gate below independently. Record for each gate:

- evidence inspected;
- observed value or behaviour;
- result: **PASS**, **FIX**, or **HOLD**;
- exact correction or exit condition when not PASS.

### G1 — Artifact identity and provenance

Verify that all inputs and outputs are uniquely identified; all recorded SHA-256 values recompute; the EEA artifact matches the canonical 2024 validated raster identity; GHSL inputs match the canonical R2023A candidate family; the FUA geometry matches the A8 contract; and the output can be traced to one commit, configuration and command.

- **PASS:** identity, lineage, version/vintage and checksums are complete and internally consistent.
- **FIX:** a bounded metadata, manifest or checksum-recording defect can be corrected without changing scientific policy.
- **HOLD:** an input cannot be authoritatively identified, retrieved or matched to its claimed bytes/geography.

### G2 — Provisional assumptions and evidence status

Verify explicit use of:

- **ASM-H05-001:** exploratory interpretation of raster values as `µg/m³`;
- **ASM-H05-002:** exclusion of `-99.0` from arithmetic without assigning it a provider-side meaning.

Confirm both remain traceable, reversible, listed as blocking gates, and attached to every dependent result. Confirm technical nodata, `-99.0`, zero and other values are not silently conflated.

- **PASS:** both assumptions are explicit, correctly bounded to RESEARCH / PROVISIONAL, and their dependent outputs are identifiable.
- **FIX:** labelling or dependency tracing is incomplete but the computation remains recoverable without a policy choice.
- **HOLD:** either assumption is hidden, broadened, treated as resolved, used outside its permitted scope, or `-99.0` is recoded/zero-filled/given an inferred provider meaning.

### G3 — Population-weighting method and temporal alignment

Verify that the estimand and weighting formula are stated; numerator and denominator populations are auditable; population weights are non-negative; zero-population handling is explicit; GHSL epoch and estimate/projection status survive the transformation; and temporal alignment is labelled as contemporaneous description, reconstructed series, or strict as-of-date hindcast as applicable.

For any 2025 projected weights or later-vintage reconstruction, confirm there is no claim of a true historical hindcast without a no-lookahead demonstration.

- **PASS:** formula, denominator, epoch/vintage, projection status and temporal claim are explicit and consistent.
- **FIX:** an implementation or labelling defect is bounded and the declared rule is otherwise unambiguous.
- **HOLD:** the temporal estimand/rule is unstated, future information is presented as an as-of-date hindcast, or weighting requires an unresolved methodology choice.

### G4 — Reprojection, resampling and boundary cells

Verify the declared source and target CRS/grid; transformation parameters; population-conserving treatment for people-per-cell counts; pollution-field treatment; grid origin/resolution; and boundary-cell rule. Confirm that coastal/ocean nodata, outside-domain cells, zero population and valid zero concentration remain distinguishable.

Require conservation checks before/after transformation and a reproducible accounting of included, excluded and fractionally allocated population.

- **PASS:** spatial operations are explicit, appropriate to variable semantics, deterministic and population accounting closes within a predeclared numerical tolerance.
- **FIX:** a coding, metadata or numerical-tolerance defect has a single bounded correction.
- **HOLD:** population conservation is not demonstrated, boundary treatment is undisclosed, variable semantics are violated, or alternative defensible treatments materially change the conclusion without a governing rule.

### G5 — Coverage and missingness

Require at minimum:

- total FUA population under the declared boundary rule;
- population with valid PM2.5 coverage;
- uncovered population and ratio;
- cells/population affected by technical nodata;
- cells/population affected by `-99.0`;
- any other non-finite/negative/excluded values;
- zero-population and outside-domain accounting;
- explicit denominator used for the reported exposure.

No missing pollution value may be zero-filled or silently dropped from the coverage denominator.

- **PASS:** cell and population coverage reconcile, exclusions are disaggregated, and the metric denominator matches the documented rule.
- **FIX:** a bounded reporting/reconciliation defect can be corrected without changing missingness policy.
- **HOLD:** material coverage is unknown, exclusions are conflated, accounting does not reconcile, or the treatment could materially bias exposure without an approved rule.

### G6 — Deterministic reproducibility

From the documented environment, rerun the fixed inputs and configuration at least twice. Compare output bytes where canonical serialization permits; otherwise compare all reported numerical and QA fields under a predeclared tolerance. Recompute all checksums.

- **PASS:** reruns reproduce the result and QA/provenance record deterministically.
- **FIX:** non-scientific nondeterminism such as timestamps or ordering has a bounded correction and does not change numerical evidence.
- **HOLD:** numerical output, population accounting, coverage or provenance changes across fixed reruns without an explained and controlled cause.

### G7 — Sensitivity and falsification

The A10 bundle must state in advance which conclusion is being tested and what change would count as material. At minimum compare:

1. preferred 100 m GHSL treatment against the retained 1 km control where both are available;
2. at least one defensible alternative alignment/boundary treatment;
3. inclusion-domain diagnostics for `-99.0` exclusion and affected-population coverage;
4. a deliberately perturbed or synthetic control capable of failing if weights, masks or denominators are wrong.

Sensitivity variants must remain labelled diagnostics, not opportunistic choices of a preferred result.

- **PASS:** tests are reproducible, capable of falsification, and the conclusion is stable under the predeclared materiality rule or explicitly qualified.
- **FIX:** a required diagnostic is missing but can be added without choosing new scientific policy.
- **HOLD:** no falsification-capable test exists, materiality is defined after seeing results, or defensible treatments lead to materially different conclusions requiring methodology adjudication.

### G8 — Comparator and sanity checks

Compare the result with:

- an unweighted valid-cell mean, clearly labelled as a diagnostic rather than the target metric;
- min/max/quantiles of included valid concentrations;
- the valid source-domain range;
- at least one independently computed aggregate or alternate implementation;
- expected arithmetic bounds: the weighted mean must be finite, non-negative, and between the minimum and maximum included concentration when covered population is positive.

Any cross-territory comparator must use the same source family, year, evidence status and transformation contract. No production ranking is permitted.

- **PASS:** arithmetic invariants hold and independent/comparator checks reveal no unexplained material discrepancy.
- **FIX:** a bounded implementation or reporting defect explains the discrepancy and is correctable.
- **HOLD:** an invariant fails, the independent computation cannot be reproduced, or a material discrepancy remains unexplained.

### G9 — MVE and promotion barrier

Run the MVE machine gate on the associated evidence package and retain its full output. Structural reviewability must not be described as scientific validation.

Because exact unit semantics and `-99.0` provider semantics remain unresolved, A10 must retain their blocking gates and remain RESEARCH / PROVISIONAL. No score, ranking, product claim, VALIDATED label or PUBLISHED claim may be derived from this review.

- **PASS:** the package is reviewable at RESEARCH status, all blockers remain explicit, and no promotion is attempted.
- **FIX:** evidence-state wording or machine-readable packaging is defective but the package remains fail-closed.
- **HOLD:** a blocking gate is removed without authoritative evidence, promotion is attempted, or public/product use overstates the evidence.

## Overall disposition

The reviewer assigns exactly one overall disposition:

### PASS

Every gate passes. The A10 artifact is reproducible and suitable to merge **only as RESEARCH / PROVISIONAL evidence**. PASS does not validate H05, resolve ASM-H05-001/002, authorize publication, or approve a production score/ranking.

### FIX

At least one gate has a precise, bounded, non-policy defect; no HOLD condition is present; and the smallest correction is identifiable. Return the package to Lane D with exact failing evidence and rerun requirements. Re-review the corrected immutable artifact.

### HOLD

Any gate has a material unresolved provenance, semantics, temporal, spatial, missingness, reproducibility, sensitivity, comparability, licence or evidence-state problem that cannot be repaired without authoritative external evidence or methodology/founder adjudication. Record the exact blocker and exit condition. Do not promote or publish.

A PASS verdict requires all gates to PASS. FIX never overrides HOLD. Uncertainty defaults to HOLD.

## Reviewer independence and closure record

The reviewer must not be the A10 producer and must not approve their own changes. The final review record must identify the reviewed artifact checksum and commit, list gate-level outcomes and limitations, and state one recommendation under the canonical Review & Closure Gate.

Until an A10 artifact exists, this protocol itself has **NO VERDICT** and creates no scientific evidence claim.